realm-lfp-retrospective-decoding

name

realm-lfp-retrospective-decoding

description

REALM methodology for LFP-based behavior decoding using retrospective distillation. Use when: building causal LFP decoding models, offline-to-online distillation for neural signals, Mamba-based neural sequence modeling, BCI decoding without spike signals, reducing bandwidth/power in implantable BCIs, behavior decoding from local field potentials. Activation: LFP decoding, REALM, retrospective distillation, causal neural decoding, wireless BCI, Mamba neural model, spike-free decoding.

REALM: Retrospective Encoder Alignment for LFP Modeling

Methodology for causal LFP-based behavior decoding via retrospective distillation.

Problem

Spike signals dominate BCI decoding but have high power/bandwidth costs. LFPs offer lower energy and bandwidth but causal LFP models show reduced accuracy vs. non-causal architectures.

Solution: REALM Framework

Transfer knowledge from offline bidirectional model to causal real-time model via retrospective distillation.

Architecture

Teacher Model (Offline)

• Backbone: Bidirectional Mamba-2
• Pretraining: Masked autoencoding objective across multiple sessions
• Purpose: Capture rich bidirectional temporal dependencies

Student Model (Real-time)

• Backbone: Compact causal Mamba-2
• Training: Combined objective of representation alignment + task supervision
• Purpose: Real-time causal decoding with minimal parameters

Distillation Pipeline

Bidirectional Teacher (offline)
    ↓ representation alignment + task supervision
Causal Student (real-time)
    ↓ behavior decoding output

Key Results

• Outperforms causal AND non-causal LFP-based SOTA methods
• Significant parameter count reduction
• Significant training time reduction
• Competitive with spike-based decoding using LFP-only signals

Workflow

1. Pretrain Teacher: Multi-session bidirectional Mamba-2 with masked autoencoding
2. Distill to Student: Align representations + task supervision
3. Deploy: Compact causal model for real-time wireless BCI

Implementation Notes

• Mamba-2 state space model captures long-range temporal dependencies efficiently
• Masked autoencoding pretraining leverages unlabeled multi-session data
• Representation alignment preserves teacher's internal representations
• Task supervision ensures decoding accuracy is maintained
• Suitable for next-generation wireless implantable BCIs

Activation Keywords

• LFP decoding
• REALM
• retrospective distillation
• causal neural decoding
• wireless BCI
• Mamba neural model
• spike-free decoding
• behavior decoding LFP
• local field potential decoding